# AI Security Institute The AI Security Institute (formerly the AI Safety Institute) is the UK government's frontier AI evaluation and safety research body — a directorate within the Department of Science, Innovation and Technology (DSIT) that conducts pre-deployment testing of advanced AI models, researches dangerous capabilities, and develops AI safety methodology. It is the leading national-government AI safety research organization globally and the UK's primary counterpart to the US [[CAISI]]. ## Origin and rebranding Established in November 2023 following the UK AI Safety Summit at Bletchley Park (hosted by then-PM Rishi Sunak), the institute was originally named the AI Safety Institute. In February 2025, the UK government rebranded it to the AI Security Institute — shifting emphasis from "safety" (associated with the alignment and ethics community) to "security" (associated with national defense and misuse prevention). The linguistic shift paralleled the US renaming of its AI Safety Institute to [[CAISI]], reflecting a transatlantic convergence on framing AI governance as a security and growth issue rather than a safety and ethics issue. ## Mission and operations The AI Security Institute's stated mission is to "minimise surprise to the UK and humanity from rapid and unexpected advances in AI." In practice, this means pre-deployment evaluation of frontier AI systems from labs including [[Anthropic]], [[OpenAI]], [[Google DeepMind]], and others — testing for dangerous capabilities in cybersecurity, biosecurity, chemical weapons, and autonomous operation. The institute has evaluated more than 30 frontier systems since its founding. The institute operates with dedicated research staff in London and maintains formal arrangements with frontier AI labs for model access. Its research agenda includes dangerous-capability evaluations, alignment methodology, AI agent safety, and systemic risk from rapid AI capability gains. ## KB relevance The AI Security Institute matters to the vault as part of the emerging transatlantic AI governance architecture. The UK's position within the [[Five Eyes Alliance]] makes its AI evaluation capabilities directly relevant to the intelligence community's assessment of AI risks — [[GCHQ]]'s cybersecurity mandate overlaps directly with the institute's evaluation of AI models' cyber capabilities. The [[Anthropic Model Export Directive]] affected Five Eyes partner nations' access to [[Claude Mythos 5]], raising the question of whether AI safety evaluation bodies like the AI Security Institute should have a role in determining which models are restricted and which are restored. ## See Also - [[CAISI]] — US counterpart (NIST Center for AI Standards and Innovation) - [[Dual-Use AI]] — the framework the institute evaluates against - [[Frontier Model]] — the class of models it tests - [[Guardrails]] — the safety mechanisms it assesses - [[GCHQ]] — UK signals intelligence with overlapping cyber mandate - [[Anthropic Model Export Directive]] — the policy domain where model evaluations matter most